nx_drbench_fetch.nx

nx_drbench_fetch.nx -- R2a: fetch a PUBLIC multi-hop QA benchmark subset sovereignly (the census [BEN] axis: "MEASURED scores on public benchmarks -- NEVER run"). Dataset = HotpotQA (Yang et al. 2018), DISTRACTOR validation split, first 100 rows, via the Hugging Face datasets-server rows API (JSON; no auth for public datasets). Distractor protocol = each row carries its own 10 context paragraphs (2 gold + 8 distractors), so retrieval+extraction is evaluable self-contained -- the published eval, not a self-authored set (the rigged-gate sin is banned). Raw JSON lands in knowledge/fetched/drb_hotpot_p0.raw for nx_drbench to parse. Composes nx_research_engine (sovereign TLS-1.3 + CA store + gzip-guard + idempotent bank). Run SOLO. expect_exit: 0 license_tier: ORIGINAL